Marcus / Felix, This looks great! Yes, this is the kind of plot I was talking about. It would also be useful to get plots for the other data in the Monitor Mixins, like # of segments, # of connected synapses per segment, etc., to get a high-level picture of how the TM is learning.
- Chetan > On Nov 30, 2015, at 1:02 PM, Marcus Lewis <[email protected]> wrote: > > Thanks for the compliments. Just to make sure the history books get it right: > I saw this thing over here, and this other thing over there, and I decided to > bring them together. All credit goes to Felix for creating the thing. It's > awesome having this tool that was shaped by tasks like implementing HTM > <https://github.com/nupic-community/comportex/>, using it > <https://nupic-community.github.io/sanity/>, and studying it > <http://floybix.github.io/>. > > Chetan: I'm glad you like it, thanks for diving in. Regarding higher-level > statistics: you, Felix, and Bill Atkinson > <https://youtu.be/OHSuydq2OW4?t=7m36s> have all thought of this. :) > > So I added it. Here's the commit > <https://github.com/nupic-community/sanity-nupic/commit/2e18722a7abb4cda5f095e205fddbbe76efb15d1>. > It was already there in the client, it just needed the HTM server to > participate. > > Here's a TM receiving predictable sequences with resets at the beginning. > Time is on the left-right axis, and number-of-columns is on the top-bottom > axis. > > <image.png> > Here's the hotgym example, using temporal_memory.py: > > <image.png> > > Here's the hotgym example, using TP10X2.py > > <image.png> > > On Mon, Nov 30, 2015 at 9:40 AM, Matthew Taylor <[email protected] > <mailto:[email protected]>> wrote: > What Marcus has done for HTM visualizations is truly amazing. If this > type of thing had existed when I first started working with HTM, the > learning curve would have been significantly decreased. > > On Wed, Nov 25, 2015 at 3:40 PM, Chetan Surpur <[email protected] > <mailto:[email protected]>> wrote: > > A nice little project for an interested community member might be to create > > an example of high-order sequence learning, visualize it with Sanity, and > > make a screencast or an interactive document that demonstrates how the > > temporal memory learns high order sequences. > > If anyone does this, I will send that person a box of goodies. :) > > --------- > Matt Taylor > OS Community Flag-Bearer > Numenta > >
